基于特征融合的车型检测新算法
New algorithm for vehicle type detection based on feature fusion
耿庆田 1于繁华 2王宇婷 3高琦坤2
作者信息
- 1. 长春师范大学计算机科学与技术学院,长春130032 ;吉林大学计算机科学与技术学院,长春130012
- 2. 长春师范大学计算机科学与技术学院,长春130032
- 3. 吉林大学计算机科学与技术学院,长春130012
- 折叠
摘要
为提高车型检测速度与精确度,本文首先通过对 HOG 特征进行改进,提出了分层HOG对称算法,得到了分层的 HOG对称特征,并将该特征与LBP特征结合,获得二者的融合特征;然后,将融合特征作为支持向量机(SVM )分类器的训练样本,采用主成分分析(PCA )法约减维数降低分类器的复杂程度;最后,使用SVM 对车辆外型特征进行检测,获得检测结果.仿真实验数据表明:该算法提高了特征提取的速度,并改善了特征检测的精度,使原始车辆图像的检测实时性得到提升;处理速度均值为26 .2帧／s ,准确率均值达到94 .58%,比传统的HOG特征提高了7 .98%.本文方法能有效提高车型检测检测的正确率,减少高维特征带来的计算时间消耗.
Abstract
To increase the speed and accuracy of vehicle recognition , an improved hierarchical Histogram of Oriented Gradient (HOG) symmetry algorithm was proposed .First ,the HOG features are improved to get the hierarchical HOG symmetry features which are fused with Local Binary Pattern (LPB ) features to get the fusion features .Second , the fusion features are taken as the training sample of Support Vector Machine (SVM ) classifier .Meanwhile ,the Principal Component Analysis (PCA ) is used to reduce the dimensions for decreasing the complexity of the classifier . Finally ,the SVM is used to recognize the appearance features of the vehicle .Simulation results show that the proposed vehicle type recognition algorithm can not only increase the feature extraction speed but also improve the detection accuracy , enhancing the real‐time recognition of original vehicle images .The mean processing speed is about 26 .2 frames／s and the accuracy is about 94 .58%, increasing by 7 .98% in comparison to traditional HOG feature extraction algorithm .The method can effectively increase the accuracy of vehicle recognition and also reduce the computing time caused by high dimensional features .
关键词
计算机应用/车型检测/HOG特征/LBP特征/特征提取/主成分分析/支持向量机Key words
computer application/vehicle type recognition/HOG feature/LBP feature/feature extraction/principal component analysis(PCA)/support vector machine(SVM)引用本文复制引用
基金项目
吉林省省级产业创新专项基金(2016C078)
吉林省产业技术研究和开发专项基金(2017C031‐2)
吉林省教育厅"十三五"科学技术研究项目(2018269)
出版年
2018
NETL
NSTL
维普
万方数据